Pipeline retargets human hand demonstrations to robot hands at 89.3% success

single source· 1 articles · confidence: medium · first seen 2026-09-22 20:00 UTC

What this means for you

Nothing to act on yet: this is a preprint, the numbers are the authors' own and have not been independently reproduced, and the abstract does not say what the linked project page releases. If you work on robot learning, the contact-aware retargeting stage is the transferable idea.

A preprint posted to arXiv on 22 September describes a three-stage pipeline that turns recordings of human hands manipulating objects into control policies for robot hands. The first stage retargets the human motion onto a different hand's geometry while preserving the demonstrated contacts; the paper reports at least an 8-point gain in contact F1 — a measure of how much demonstrated contact is reproduced — over the best of five baselines. Residual reinforcement learning then refines the motion for physical feasibility, and the result is distilled into a visuomotor policy. Those policies reached 89.3% success in 300 real-world trials on 30 objects, the authors' own figures.

Key facts

  • ·The paper was posted to arXiv as 2609.28660 on 22 September 2026. source
  • ·The framework has three stages: morphometric optimisation, residual reinforcement learning, and distillation into visuomotor policies. source
  • ·Morphometric optimisation raises contact F1 by at least 8 points over the strongest of five baselines, for each of three robot hands. source
  • ·The same stage improves the success rate of downstream dynamic retargeting by as much as 35 points, per the paper's ablations. source
  • ·The resulting visuomotor policies reached 89.3% zero-shot success across 300 real-world trials on 30 objects. source
  • ·Evaluation covered three robot hands and ten human hand-object interactions. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire